CN116127201A - Large-scale user recommendation method based on evolutionary multitasking - Google Patents
Large-scale user recommendation method based on evolutionary multitasking Download PDFInfo
- Publication number
- CN116127201A CN116127201A CN202310167737.1A CN202310167737A CN116127201A CN 116127201 A CN116127201 A CN 116127201A CN 202310167737 A CN202310167737 A CN 202310167737A CN 116127201 A CN116127201 A CN 116127201A
- Authority
- CN
- China
- Prior art keywords
- user
- generation
- recommended
- recommendation
- article
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 48
- 230000003993 interaction Effects 0.000 claims abstract description 38
- 238000013508 migration Methods 0.000 claims abstract description 16
- 230000005012 migration Effects 0.000 claims abstract description 16
- 238000005457 optimization Methods 0.000 claims abstract description 14
- 239000011159 matrix material Substances 0.000 claims abstract description 11
- 239000013598 vector Substances 0.000 claims description 32
- 230000035772 mutation Effects 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 6
- 238000013507 mapping Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 238000004590 computer program Methods 0.000 claims description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000002860 competitive effect Effects 0.000 claims description 3
- 230000013011 mating Effects 0.000 claims description 3
- 238000010187 selection method Methods 0.000 claims description 3
- 230000002452 interceptive effect Effects 0.000 claims description 2
- 125000003275 alpha amino acid group Chemical group 0.000 claims 1
- 150000001875 compounds Chemical class 0.000 claims 1
- 238000013528 artificial neural network Methods 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 3
- 210000000349 chromosome Anatomy 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9536—Search customisation based on social or collaborative filtering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Genetics & Genomics (AREA)
- Artificial Intelligence (AREA)
- Physiology (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a large-scale user recommendation method based on evolutionary multitasking, which comprises the following steps: step 1, acquiring a user and article interaction data set, and step 2, excavating preference scores of the user on articles by constructing a neural network algorithm to obtain a scoring matrix of the user on the articles; step 3: grouping users through clustering, and taking users with similar preference interests as the same category; step 4, initializing and generating a multi-task population; and 5, performing information migration among individuals in the same user group, performing information migration among populations among different user groups, iteratively selecting an optimal user solution through environment selection, and finally selecting an optimal solution as a recommendation list of the user. The invention can reduce the time and space consumed in large-scale user recommendation optimization problem, and improves the accuracy of predicting the user recommendation result through the clustering technology.
Description
Technical Field
The invention belongs to the crossing field of evolutionary computation and data mining, and particularly relates to a recommendation method based on evolutionary multitasking optimization.
Background
The aim of the recommendation system is to help users to screen useful information from massive data, the traditional recommendation system only considers the recommendation accuracy, and besides the accuracy, other performances such as novelty, diversity and the like are also important indexes of the recommendation system, so that the multi-target recommendation system becomes an important research direction of the recommendation system. However, as the recommendation system needs to meet multiple indexes, optimizing some indexes inevitably brings conflicts to some indexes. Therefore, multi-objective optimization becomes an important technical means for solving the recommendation system.
The existing multi-objective optimization method of the recommendation system comprises the following steps: the multi-objective index is weighted and summed into a scalar method of a single objective problem, or a population-based evolutionary algorithm is used to simultaneously optimize multiple objectives.
Scalar methods first typically sum two target weights and then combine the scalar with pareto high efficiency SGD using a multiple gradient descent algorithm, using KKT conditions to direct the updating of scalar weights. However, this method can only be optimized for targets with gradients.
When the evolutionary algorithm optimizes the multi-objective recommendation system, the evolutionary algorithm is usually operated for each user respectively, however, the evolutionary algorithm is used as an iterative algorithm, and when the number of users in the data set is excessive, the algorithm is sequentially optimized, so that the operation time of the algorithm is excessively long. Or the recommended results of the users are borrowed into one chromosome through real solution, and all users are simultaneously optimized in one optimization process, but the method can lead to overlong encoding length of the chromosome, and is difficult to simultaneously optimize all users to be optimal, so that the optimization method is easy to fall into a local optimal solution, and the recommended article results can not meet the user requirements.
Disclosure of Invention
The invention provides a large-scale user recommending method based on evolutionary multitasking for solving the defects in the prior art, thereby realizing a recommending list containing accuracy, novelty and diversity for users and simultaneously ensuring the high efficiency and the sum of the recommending methods.
The invention adopts the following technical scheme for solving the technical problems:
the invention discloses a large-scale user recommendation method based on evolutionary multitasking, which is characterized by comprising the following steps:
step one, acquiring related data of a user and an article:
acquiring a user set s= { S 1 ,s 2 ,…,s u ,…,s |S| -wherein S represents the number of users S u The user u is represented;
acquiring an item set q= { Q 1 ,q 2 ,…,q i ,…,q |Q| -wherein Q represents the number of items, Q i Representing an ith item;
acquiring a user article interaction data set, taking interaction data of each user and articles as positive samples, and randomly collecting articles which are not interacted by the user as negative samples;
step two, obtaining a scoring matrix of the user on the article through a prediction model, wherein the prediction model comprises the following steps: the system comprises a coding layer, a full connection layer, an attention interaction layer, an interaction output layer and a prediction layer;
step 2.1, the coding layer is used for the u number user s u And the ith article q i Performing one-hot coding to obtain u-number user s u Is a sparse vector theta of (2) u And article q i Is a sparse vector of (2)Then mapping the two sparse vectors to E-dimensional vectors respectively, and obtaining u-number user s through mapping of the full connection layer u Is a representation vector p of (2) u And the ith article q i Is a representation vector o of (2) i ;
Step 2.2, the attention interaction layer pair p u And o i Processing and outputting attention vector a u,i ;
Step 2.3, the interactive output layer will p u And o i After splicing, the vector is combined with the attention vector a u,i After dot multiplication, obtaining a u-number user s u And the ith article q i Is of the interaction vector f u,i ;
Step 2.4, the prediction layer pair interaction vector f u,i After multi-layer full-connection processing, a predicted interaction score r is output u,i ;
Step 2.5, taking the mean square error of the minimized predicted interaction score as a loss function, and optimizing the prediction model by using an Adam algorithm until the maximum iteration number is reached, so as to obtain an optimal prediction model and an interaction score matrix of each item predicted by each user output by the optimal prediction model;
step three, setting the clustering number as K, wherein top represents the number of candidate articles, and selecting u-number users s u Top preferred items of (a) as candidate item set candidates u Counting the number of the same articles in the candidate article sets of each user as the similarity between the corresponding users; users with similar similarity are used as a class through a clustering algorithm, so that the users are divided into K groups by S|, and a user set U= { U is obtained 1 ,U 2 ,…,U j ,…,U K },U j User group representing a j-th group category, and U j ={P j,1 ,P j,2 ,…,P j,m ,…,P j,M };P j,m Representing U j P-th user of (a);
step four, initializing a population:
step 4.1, defining the current iteration number as L, the maximum iteration number as L, enabling N to be the population individual number, and adopting a real number system to enable the user group U of the j group category to be the j group j N recommendation results of each user are respectively encoded into an individual with the length of T, and each decision variable of the individual represents the serial number of the recommended article; thus, N recommendation results of one user form a group, and the mth user P is led to j,m N recommended results of (a) are recorded as the mth species of the first generationGroup, user group U of j-th group category j All first generation populations in (a) are marked as
U is set to j M th user P j,m The nth recommendation of (2) is recorded as the nth individual in the mth population of the first generationAnd is also provided with Representing the mth generation of users P j,m The number of the t recommended article in the nth recommended result;
step 4.2, from the mth user P, based on the interaction score matrix predicted by each user for each item j,m Candidate item set of (5) j,m Randomly selecting the serial numbers of T non-repeated articles for pairingInitializing;
step 4.3, obtaining the mth generation of the mth user P by using the formula (1) j,m Is the nth recommendation result of (2)Accuracy index>
In the formula (1), the components are as follows,representing the mth user P j,m For serial number->Scoring the corresponding item;
obtaining the mth generation user P by using the method (2) j,m Is the nth recommendation result of (2)Novel rate index of (a)/>
obtaining the mth generation user P by using the method (3) j,m Is the nth recommendation result of (2)Is a diversity index of (2)
In the formula (3), labelIndicating number->Category label of corresponding article, label all Representing user item interaction data set propertiesCategory labels of the products;
constructing a first generation multi-objective optimization function maxisize by using the formula (4)
And 5, performing information migration among individuals among the same user group, performing information migration among populations among different user groups, and iteratively selecting an optimal user solution through environment selection.
The large-scale user recommendation method based on evolutionary multitasking of the invention is also characterized in that the step 5 comprises:
step 5.1, information migration among individuals is carried out among the same user group:
step 5.1.1 Using binary tournament selection method based on equation (4)Selecting 2 XN recommended results to participate in evolution to obtain first generation mating pool ∈>
Step 5.1.2 from the first Generation poolSelecting two recommended results of the first generation and marking the recommended results as +.> And->And performing cross operation to obtain two first generationCross recommendation result->Wherein (1)>Representation->The number of the t-th recommended item,/-)>Representation->The number of the t recommended article;
randomly selecting a number r' from {1,2,3 …, T }, fromCandidate item sets corresponding to affiliated users j,1 Is selected randomly from one and->The order of the r' th recommended item +.>The different article serial numbers are replaced, thereby obtaining +.>Recommended first generation variant of ++>
Step 5.1.4 for the first Generation poolAfter crossing and mutation operations are carried out on all recommended results in the first generation of all populations according to the steps 5.1.2-5.1.3, mutation recommended results of all populations in the first generation are obtained>Will->And->User group U combined into j-th group category j The method comprises the steps of (1) measuring the fitness value of each recommended result in the first-generation M new populations through a formula (4), so that environment selection is carried out on the first-generation M new populations through non-dominant sorting and crowding distances, and the optimal N recommended results are reserved as the first-generation (1) M populations;
step 5.2, information migration among the populations is carried out among different user groups:
if the mth generation of users P j,m If more than half of the recommended results are unchanged for a plurality of successive generations, computing a user group U of the j-th group category j Similarity between user groups of other group categories, and selecting the user group with highest similarity for the mth user P of the first generation j,m Performing crossover and mutation operations between all recommended results of (1) so as to obtain an mth population of the (1+1) th generation according to the process of the step 5.1.4;
step 5.3, after assigning l+1 to L, judging whether L reaches L, if not, returning to step 4.3 for sequential execution, otherwise, selecting an individual from the mth population of the L generation as a user group U of the j-th group class j M th user P j,m Is a recommended result of the user.
The step 5.1.2 comprises:
step a, judgingAnd->If the users belong to the same user, executing the step b, otherwise, executing the step c;
randomly selecting a number r from {1,2,3 …, T }, willThe former r position and->The first r bits of (a) are exchanged to obtain the first generation two cross recommended results +.>And-> Wherein (1)>Representation->The order of the r-th recommended item,/-)>Representation ofMiddle (f)The serial numbers of r recommended articles;
step c, willAnd->Combining to obtain a first-generation new recommended result, using formula (1) as fitness value, using binary competitive competition method to respectively select serial numbers corresponding to T items from the first-generation new recommended result and respectively forming two first-generation cross recommended results->To->As->Cross-recommendation results generated by +.>As->The cross recommendation results are generated.
The electronic device of the invention comprises a memory and a processor, wherein the memory is used for storing a program for supporting the processor to execute any large-scale user recommendation method, and the processor is configured to execute the program stored in the memory.
The invention provides a computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to execute the steps of any large-scale user recommendation method.
Compared with the prior art, the invention has the beneficial effects that:
1. according to the method, the evolutionary multitasking is used for optimizing the multi-target recommendation system for the first time, the traditional optimization method can only generate an optimized recommendation result for one user at a time, and correlation among different tasks is not considered. According to the method and the device, the optimal recommendation results of a plurality of users can be simultaneously generated in one optimization process, so that the calculated amount of a recommendation algorithm is greatly reduced.
2. In order to avoid the negative migration phenomenon possibly caused by simultaneous optimization, the invention uses a clustering algorithm to take a plurality of users with similar interests in data as a user group in an optimization method, and designs a new operator to accelerate the convergence of the algorithm, thereby improving the recommendation speed.
3. According to the invention, by designing a population optimization scheme, genetic operators and information migration strategies of different populations, the calculated amount is reduced, the recommendation efficiency is improved, and the recommendation accuracy is ensured.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a diagram illustrating an example of user clustering in accordance with the present invention;
FIG. 3 is an exemplary diagram of population initialization in accordance with the present invention;
FIG. 4 is a diagram illustrating the generation of a next generation population according to the present invention.
Detailed Description
In this embodiment, a large-scale user recommendation method based on evolutionary multitasking, as shown in fig. 1, is performed according to the following steps:
step one, acquiring related data of a user and an article:
acquiring a user set s= { S 1 ,s 2 ,…,s u ,…,s |S| -wherein S represents the number of users S u The user u is represented;
acquiring an item set q= { Q 1 ,q 2 ,…,q i ,…,q |Q| -wherein Q represents the number of items, Q i Representing an ith item;
acquiring a user article interaction data set, taking interaction data of each user and articles as positive samples, and randomly collecting articles which are not interacted by the user as negative samples;
step two, obtaining a scoring matrix of the user on the article through a prediction model, wherein the prediction model comprises: the system comprises a coding layer, a full connection layer, an attention interaction layer, an interaction output layer and a prediction layer;
step 2.1, coding layer pair u number user s u And the ith article q i Performing one-hot coding to obtain u-number user s u Is a sparse vector theta of (2) u And article q i Is a sparse vector of (2)Then mapping the two sparse vectors to E-dimensional vectors respectively, and obtaining u-number user s through mapping of a full connection layer u Is a representation vector p of (2) u And the ith article q i Is a representation vector o of (2) i ;
Step 2.2, attention interaction layer pair p u And o i Processing and outputting attention vector a u,i ;
Step 2.3, the interaction output layer outputs p u And o i After splicing, the vector is combined with the attention vector a u,i After dot multiplication, obtaining a u-number user s u And the ith article q i Is of the interaction vector f u,i ;
Step 2.4, prediction layer pair interaction vector f u,i After multi-layer full-connection processing, a predicted interaction score r is output u,i ;
Step 2.5, taking the mean square error of the interaction score of the minimum prediction as a loss function, and optimizing the prediction model by using an Adam algorithm until the maximum iteration number is reached, so as to obtain an optimal prediction model and an interaction score matrix of each user output by the optimal prediction model for each article;
step three, setting the clustering number as K, wherein top represents the number of candidate articles, and selecting u-number users s u Top preferred items of (a) as candidate item set candidates u Counting the number of the same articles in the candidate article sets of each user as the similarity between the corresponding users; users with similar similarity are used as a class through a clustering algorithm, so that the users are classified into K groups by S, and in the embodiment, the method comprises the following steps ofK= |s|/10 and a user set u= { U is obtained 1 ,U 2 ,…,U j ,…,U K },U j User group representing a j-th group category, and U j ={P j,1 ,P j,2 ,…,P j,m ,…,P j,M };P j,m Representing U j P-th user of (a); as in the example of fig. 2, it is assumed that there are 4 users s 1 ,s 2 ,s 3 ,s 4 The candidate item set of (2) is obtained by counting the same item number of different users 1 Sum s 2 Is 7,s 3 Sum s 4 The similarity of (2) is 7, so the final clustering result is s 1 Sum s 2 Is U (U) 1 Class, s 3 Sum s 4 Is U (U) 2 Class.
Step four, initializing a population:
step 4.1, defining the current iteration number as L and the maximum iteration number as L, in this embodiment, setting l=100, and making N be the population individual number, in this embodiment, setting n=10, and using real number system to make user group U of the j-th group category j N recommendation results of each user are respectively encoded into an individual with a length of T, in this embodiment, t=10 is set, and each decision variable of the individual represents a serial number of a recommended article; thus, N recommendation results of one user form a group, and the mth user P is led to j,m The N recommended results of the (a) are recorded as the mth population of the first generation, and the user group U of the jth group category j All first generation populations in (a) are marked as
U is set to j M th user P j,m The nth recommendation of (2) is recorded as the nth individual in the mth population of the first generationAnd is also provided with Representing the mth generation of users P j,m The nth recommended article serial number in the nth recommended result, the population individuals are composed of the recommended results of all users of the group, and the recommended results are encoded into a matrix. Fig. 3 shows an example of a class of population individuals with a number of users of 2 and a recommended length of 10.
Step 4.2, from the mth user P, based on the interaction score matrix predicted by each user for each item j,m Candidate item set of (5) j,m Randomly selecting the serial numbers of T non-repeated articles for pairingInitializing;
step 4.3, obtaining the mth generation of the mth user P by using the formula (1) j,m Is the nth recommendation result of (2)Accuracy index>
In the formula (1), the components are as follows,representing the mth user P j,m For->Scoring of the item;
obtaining the mth generation user P by using the method (2) j,m Is the nth recommendation result of (2)Novel rate index of (a)
obtaining the mth generation user P by using the method (3) j,m Is the nth recommendation result of (2)Is a diversity index of (2)
In the formula (3), labelRepresentation->Category label of article, label all Category labels representing all items in the user item interaction dataset;
constructing a first generation multi-objective optimization function maxisize by using the formula (4)
step 5.1, information migration among individuals is carried out among the same user group:
step 5.1.1 Using binary tournament selection method based on equation (4)Selecting 2 XN recommended results to participate in evolution to obtain first generation mating pool ∈>
Step 5.1.2 from the first Generation poolSelecting two recommended results of the first generation and marking the recommended results as +.> And->And performing a crossover operation wherein->Representation->The number of the t-th recommended item,/-)>Representation->The number of the t recommended item:
step a, judgingAnd->If the users belong to the same user, executing the step b, otherwise, executing the step c;
randomly selecting a number r from {1,2,3 …, T }, willThe former r position and->The first r bits of (a) are exchanged to obtain the first generation two cross recommended results +.>And-> Wherein (1)>Representation->The order of the r-th recommended item,/-)>Representation ofThe number of the r-th recommended article;
step c, willAnd->Combining to obtain a first-generation new recommended result, using formula (1) as fitness value, using binary competitive competition method to respectively select serial numbers corresponding to T items from the first-generation new recommended result and respectively forming two first-generation cross recommended results->To->As->Cross-recommendation results generated by +.>As->Generating a cross recommendation result;
randomly selecting a number r' from {1,2,3 …, T }, fromCandidate item sets corresponding to affiliated users j,1 Is selected randomly from one and->The r' th recommended article in the list/>The different article serial numbers are replaced, thereby obtaining +.>Recommended first generation variant of ++>
Fig. 4 shows an example illustrating the specific operation of the crossover and mutation operator. The class of user population in the example includes 2 users, denoted s 1 ,s 2 ,X 1 ,X 2 Is s 1 ,s 2 Is 10, probability c P ,m P 0.5 and 0.5, respectively. First, the crossover operation is performed, and assuming that the generated random number is 0.3, then for X 1 ,X 2 Cross operation is performed on X 1 Randomly selecting a number from {1,2,..10 }, and if 3, crossing (1,23,15,9,5,12,4,18,22,14), (18,2,16,8,20,24,4,25,17,30) to obtain (1,23,15,8,20,24,4,25,17,30), (18,2,16,9,5,12,4,18,22,14). Let the same assumption X 2 Randomly selecting one number, which is 4, (10,2,7,6,5,3,11,13,17,19), (10,2,15,6,21,12,11,27,28,29) to obtain (10,2,7,6,21,12,11,27,28,29), (10,2,15,6,5,3,11,13,17,19) after crossing, to obtain two childrenThen, a mutation operator is performed for each child. For->Assuming that the corresponding generated random number is 0.4, then for +.>A mutation operation was performed, and a number selected from {1,2,..10 } was randomly selected, assuming 4. From user s 1 Candidate item set candidates of (c) s1 A new article 6 is selected to replace the original article 8. Pair s of the same theory 2 Performing similar operations, finally->Obtaining->They are two new recommendations.
Step 5.1.4 for the first Generation poolAfter crossing and mutation operations are carried out on all recommended results in the first generation of all populations according to the steps 5.1.2-5.1.3, mutation recommended results of all populations in the first generation are obtained>Will->And->User group U combined into j-th group category j The method comprises the steps of (1) measuring the fitness value of each recommended result in the first-generation M new populations through a formula (4), so that environment selection is carried out on the first-generation M new populations through non-dominant sorting and crowding distances, and the optimal N recommended results are reserved as the first-generation (1) M populations;
step 5.2, information migration among the populations is carried out among different user groups:
if the mth generation of users P j,m If more than half of the recommended results are unchanged for a plurality of successive generations, computing a user group U of the j-th group category j Similarity between user groups of other group categories, and selecting the user group with highest similarity for the mth user P of the first generation j,m Performing crossover and mutation operations between all recommended results of (1) so as to obtain an mth population of the (1+1) th generation according to the process of the step 5.1.4;
step 5.3, after assigning l+1 to L, judging whether L reaches L, if not, returning to step 4.3 for sequential execution, otherwise, selecting an individual from the mth population of the L generation as a user group U of the j-th group class j M th user P j,m Is a recommended result of the user.
In this embodiment, an electronic device includes a memory for storing a program supporting the processor to execute the above method, and a processor configured to execute the program stored in the memory.
In this embodiment, a computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the method described above.
Claims (5)
1. The large-scale user recommendation method based on the evolutionary multitasking is characterized by comprising the following steps of:
step one, acquiring related data of a user and an article:
acquiring a user set s= { S 1 ,s 2 ,…,s u ,…,s |S| -wherein S represents the number of users S u The user u is represented;
acquiring an item set q= { Q 1 ,q 2 ,…,q i ,…,q |Q| -wherein Q represents the number of items, Q i Representing an ith item;
acquiring a user article interaction data set, taking interaction data of each user and articles as positive samples, and randomly collecting articles which are not interacted by the user as negative samples;
step two, obtaining a scoring matrix of the user on the article through a prediction model, wherein the prediction model comprises the following steps: the system comprises a coding layer, a full connection layer, an attention interaction layer, an interaction output layer and a prediction layer;
step 2.1, the coding layer is used for the u number user s u And the ith article q i Performing one-hot coding to obtain u-number user s u Is a sparse vector theta of (2) u And article q i Is sparse toMeasuring amountThen mapping the two sparse vectors to E-dimensional vectors respectively, and obtaining u-number user s through mapping of the full connection layer u Is a representation vector p of (2) u And the ith article q i Is a representation vector o of (2) i ;
Step 2.2, the attention interaction layer pair p u And o i Processing and outputting attention vector a u,i ;
Step 2.3, the interactive output layer will p u And o i After splicing, the vector is combined with the attention vector a u,i After dot multiplication, obtaining a u-number user s u And the ith article q i Is of the interaction vector f u,i ;
Step 2.4, the prediction layer pair interaction vector f u,i After multi-layer full-connection processing, a predicted interaction score r is output u,i ;
Step 2.5, taking the mean square error of the minimized predicted interaction score as a loss function, and optimizing the prediction model by using an Adam algorithm until the maximum iteration number is reached, so as to obtain an optimal prediction model and an interaction score matrix of each item predicted by each user output by the optimal prediction model;
step three, setting the clustering number as K, wherein top represents the number of candidate articles, and selecting u-number users s u Top preferred items of (a) as candidate item set candidates u Counting the number of the same articles in the candidate article sets of each user as the similarity between the corresponding users; users with similar similarity are used as a class through a clustering algorithm, so that the users are divided into K groups by S|, and a user set U= { U is obtained 1 ,U 2 ,…,U j ,…,U K },U j User group representing a j-th group category, and U j ={P j,1 ,P j,2 ,…,P j,m ,…,P j,M };P j,m Representing U j P-th user of (a);
step four, initializing a population:
step 4.1,Defining the current iteration number as L, the maximum iteration number as L, enabling N to be the population individual number, and adopting a real number system to set the user group U of the j group category j N recommendation results of each user are respectively encoded into an individual with the length of T, and each decision variable of the individual represents the serial number of the recommended article; thus, N recommendation results of one user form a group, and the mth user P is led to j,m The N recommended results of the (a) are recorded as the mth population of the first generation, and the user group U of the jth group category j All first generation populations in (a) are marked as
U is set to j M th user P j,m The nth recommendation of (2) is recorded as the nth individual in the mth population of the first generationAnd is also provided with Representing the mth generation of users P j,m The number of the t recommended article in the nth recommended result;
step 4.2, from the mth user P, based on the interaction score matrix predicted by each user for each item j,m Candidate item set of (5) j,m Randomly selecting the serial numbers of T non-repeated articles for pairingInitializing;
step 4.3, obtaining the mth generation of the mth user P by using the formula (1) j,m Is the nth recommendation result of (2)Accuracy index>
In the formula (1), the components are as follows,representing the mth user P j,m For serial number->Scoring the corresponding item;
obtaining the mth generation user P by using the method (2) j,m Is the nth recommendation result of (2)Novel index->
obtaining the mth generation user P by using the method (3) j,m Is the nth recommendation result of (2)Diversity index->
In the formula (3), the amino acid sequence of the compound,indicating the serial number x j l ,m,n,t Category label of corresponding article, label all Category labels representing all items in the user item interaction dataset;
And 5, performing information migration among individuals among the same user group, performing information migration among populations among different user groups, and iteratively selecting an optimal user solution through environment selection.
2. The evolutionarily multitasking-based large-scale user recommendation method of claim 1, wherein said step 5 comprises:
step 5.1, information migration among individuals is carried out among the same user group:
step 5.1.1 Using binary tournament selection method based on equation (4)Selecting 2 XN recommended results to participate in evolution to obtain first generation mating pool ∈>
Step 5.1.2 from the first Generation poolSelecting two recommended results of the first generation and marking the recommended results as +.> And->And performing cross operation to obtain two first generation cross recommendation results +.>Wherein (1)>Representation->The number of the t-th recommended item,/-)>Representation->The number of the t recommended article;
randomly selecting a number r from {1,2,3 …, T } ′ From the slaveCandidate item sets corresponding to affiliated users j,1 Is selected randomly from one and->Middle (r) ′ Number ∈of each recommended item>The different article serial numbers are replaced, thereby obtaining +.>Recommended first generation variant of ++>
Step 5.1.4 for the first Generation poolAfter crossing and mutation operations are carried out on all recommended results in the first generation of all populations according to the steps 5.1.2-5.1.3, mutation recommended results of all populations in the first generation are obtained>Will->And->User group U combined into j-th group category j The method comprises the steps of (1) measuring the fitness value of each recommended result in the first-generation M new populations through a formula (4), so that environment selection is carried out on the first-generation M new populations through non-dominant sorting and crowding distances, and the optimal N recommended results are reserved as the first-generation (1) M populations;
step 5.2, information migration among the populations is carried out among different user groups:
if the mth generation of users P j,m If more than half of the recommended results are unchanged for a plurality of successive generations, computing a user group U of the j-th group category j Similarity between user groups of other group categories, and selecting the user group with highest similarity for the mth user P of the first generation j,m Performing crossover and mutation operations between all recommended results of (1) so as to obtain an mth population of the (1+1) th generation according to the process of the step 5.1.4;
step 5.3, after assigning l+1 to L, judging whether L reaches L, if not, returning to step 4.3 for sequential execution, otherwise, selecting an individual from the mth population of the L generation as the j-th group classUser group U of (2) j M th user P j,m Is a recommended result of the user.
3. The evolutionarily multitasking-based large-scale user recommendation method of claim 2, wherein said step 5.1.2 comprises:
step a, judgingAnd->If the users belong to the same user, executing the step b, otherwise, executing the step c;
randomly selecting a number r from {1,2,3 …, T }, willThe former r position and->The first r bits of (a) are exchanged to obtain the first generation two cross recommended results +.>And-> Wherein (1)>Representation->The order of the r-th recommended item,/-)>Representation ofThe number of the r-th recommended article;
step c, willAnd->Combining to obtain a first-generation new recommended result, using formula (1) as fitness value, using binary competitive competition method to respectively select serial numbers corresponding to T items from the first-generation new recommended result and respectively forming two first-generation cross recommended results->To->As->Cross-recommendation results generated by +.>As->The cross recommendation results are generated.
4. An electronic device comprising a memory and a processor, wherein the memory is for storing a program supporting the processor to perform the large-scale user recommendation method of any one of claims 1-3, the processor being configured to execute the program stored in the memory.
5. A computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor performs the steps of the large-scale user recommendation method according to any of claims 1-3.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310167737.1A CN116127201A (en) | 2023-02-27 | 2023-02-27 | Large-scale user recommendation method based on evolutionary multitasking |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310167737.1A CN116127201A (en) | 2023-02-27 | 2023-02-27 | Large-scale user recommendation method based on evolutionary multitasking |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116127201A true CN116127201A (en) | 2023-05-16 |
Family
ID=86310037
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310167737.1A Pending CN116127201A (en) | 2023-02-27 | 2023-02-27 | Large-scale user recommendation method based on evolutionary multitasking |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116127201A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116596640A (en) * | 2023-07-19 | 2023-08-15 | 国网山东省电力公司营销服务中心(计量中心) | Recommendation method, system, equipment and storage medium for electric power retail electric charge package |
-
2023
- 2023-02-27 CN CN202310167737.1A patent/CN116127201A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116596640A (en) * | 2023-07-19 | 2023-08-15 | 国网山东省电力公司营销服务中心(计量中心) | Recommendation method, system, equipment and storage medium for electric power retail electric charge package |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110097185B (en) | Optimization model method based on generation of countermeasure network and application | |
Zhou et al. | A survey on evolutionary construction of deep neural networks | |
US20190180187A1 (en) | Evolving Recurrent Networks Using Genetic Programming | |
CN112395487B (en) | Information recommendation method and device, computer readable storage medium and electronic equipment | |
CN116127201A (en) | Large-scale user recommendation method based on evolutionary multitasking | |
Bai et al. | A joint multiobjective optimization of feature selection and classifier design for high-dimensional data classification | |
CN112836794B (en) | Method, device, equipment and storage medium for determining image neural architecture | |
CN116340726A (en) | Energy economy big data cleaning method, system, equipment and storage medium | |
CN111325284A (en) | Self-adaptive learning method and device based on multi-target dynamic distribution | |
Hussein et al. | Deep learning and machine learning via a genetic algorithm to classify breast cancer DNA data | |
Mo et al. | Applications of machine learning in phylogenetics | |
Dutta et al. | Clustering by multi objective genetic algorithm | |
CN116611504A (en) | Neural architecture searching method based on evolution | |
Jia et al. | Weak-restriction bi-objective optimization algorithm for scheduling with rejection on non-identical batch processing machines | |
Sanchez | Reconstructing our past˸ deep learning for population genetics | |
KR20230170680A (en) | Multi-channel protein voxelization to predict variant pathogenicity using deep convolutional neural networks | |
CN110162704B (en) | Multi-scale key user extraction method based on multi-factor genetic algorithm | |
Sharma et al. | Prediction of cardiovascular diseases using genetic algorithm and deep learning techniques | |
CN114443986A (en) | Sorting method and device, sorting model training method and device, and electronic equipment | |
CN113704570A (en) | Large-scale complex network community detection method based on self-supervision learning type evolution | |
WO2023044927A1 (en) | Rna-protein interaction prediction method and apparatus, and medium and electronic device | |
Hamazaki et al. | AI-assisted selection of mating pairs through simulation-based optimized progeny allocation strategies in plant breeding | |
CN117976047B (en) | Key protein prediction method based on deep learning | |
Zelený et al. | Multi-Branch Multi Layer Perceptron: A Solution for Precise Regression using Machine Learning | |
CN114220127B (en) | Image recognition method based on gradient guided evolutionary algorithm |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |